Immunity to Device Variations in a Spiking Neural Network With Memristive Nanodevices
read more
Citations
Unsupervised learning of digit recognition using spike-timing-dependent plasticity.
Neuromorphic computing using non-volatile memory
Deep learning in spiking neural networks
Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing
A Survey of Neuromorphic Computing and Neural Networks in Hardware.
References
Gradient-based learning applied to document recognition
The missing memristor found
Nanoscale Memristor Device as Synapse in Neuromorphic Systems
Regulation of Synaptic Efficacy by Coincidence of Postsynaptic APs and EPSPs
‘Memristive’ switches enable ‘stateful’ logic operations via material implication
Related Papers (5)
A million spiking-neuron integrated circuit with a scalable communication network and interface
Frequently Asked Questions (12)
Q2. What future works have the authors mentioned in the paper "Immunity to device variations in a spiking neural network with memristive nanodevices" ?
Future work should focus on the experimental demonstration of these concepts beyond single devices, and to demonstrate its scaling to more complex multi-layer networks, and to other kinds of sensory stimuli like video, auditory or olfactory data.
Q3. What is the common approach used in neuromorphic design?
An approach widely studied in the neuromorphic community is to use analog circuits (generally with transistors operating in the sub-threshold regime) able to receive and generate asynchronous spikes [6], [7] to design such neurons.
Q4. What is the way to improve the recognition rate of the network?
To improve the recognition rate, the authors can introduce additional output neurons, in which case some output neurons respond to different handwritings of the same digit.
Q5. What is the basic principle of the learning rule of the nanodevices?
they adapt their conductance depending on the activity of the neurons to which they are connected, which provides the foundation of learning by the system.
Q6. What is the reason why the CMOS network is designed to limit the sneak path issue?
Of particular interest, in the case of their network, it limits the sneak path issue because both programming and reading are performed in parallel.
Q7. What is the effect of the readdisturb parameter on the device?
It appears that readdisturb parameter as high as 0.1 (meaning that a read pulse has 10% of the impact of a write pulse) may be tolerated, since it is fully compensated by learning.
Q8. What is the effect of the variability on the recognition rate of the network?
With an extreme variability of 100% on the synaptic parameters, the recognition rate decreases significantly, but interestingly the functionality of the network is not challenged.
Q9. How many output neurons are used to associate the digits?
For that purpose, the authors associate output neurons with the digit for which they spike the most frequently a posteriori, using a subset of 1000 well identified numbers.
Q10. What is the effect of the voltage applied on the device?
When the voltage applied on the device (difference between the voltages applied at the two ends (c) or (d)) reaches VT+ or VT−, its conductance is increased or decreased, respectively.
Q11. What is the strongest point of the approach?
This degree of robustness to device variation is exceptional in electronic systems and constitutes one of the strongest points of the approach.
Q12. How can a memristive device perform unsupervised learning?
In this work, using system-level simulations, the authors have shown how, by using a simple custom Spike Timing Dependent Plasticity scheme, memristive devices associated with CMOS neuromorphic circuits could perform unsupervised learning in a way that is extremely robust to variability.